A modular neural network architecture with approximation capability and its applications

Author

Cai, Changlin ; Shi, Zhongzhi

Author_Institution

Inst. of Comput. Technol., Chinese Acad. of Sci., Beijing, China

fYear

2003

fDate

18-20 Aug. 2003

Firstpage

60

Lastpage

64

Abstract

In this paper a new modular architecture of neural networks is designed to show that any continuous function which defined on a compact set can be approximated by a multilayer perceptrons, when the output layer activation functions are linear, and the hidden layer activation function could be chosen in the conditions of no bounded and no sigmoid. An application in econometrics forecast is proposed and analyzed, where the new function models can be added to the system one by one, so the complex system can be formed easily.

Keywords

econometrics; multilayer perceptrons; neural net architecture; approximation capability; continuous function; econometrics forecast; hidden layer activation function; modular architecture; multilayer perceptrons; neural networks; no bounded condition; no sigmoid condition; output layer activation functions; Biological neural networks; Computer architecture; Computer networks; Econometrics; Economic forecasting; Multi-layer neural network; Multilayer perceptrons; Neural networks; Neurons; Nonhomogeneous media;

fLanguage

English

Publisher

ieee

Conference_Titel

Cognitive Informatics, 2003. Proceedings. The Second IEEE International Conference on

Print_ISBN

0-7695-1986-5

Type

conf

DOI

10.1109/COGINF.2003.1225954

Filename

1225954